[The original does not have parts, but I have divided it in two for presentation purposes.]

Transcription

1 Derivation by Phase Noam Chomsky GLOSS: As I did before, I ll try to gloss Chomsky s recent paper, to the extent I understand it. Please do not circulate this commentary without permission, or Chomsky s paper in this format, since although he has given permission for the commentary, the paper is to appear in a festschrift for Ken Hale. In fact, the electronic manuscript I have is the one Chomsky sent, which is slightly less worked out than the version circulated in MITWPL (M-version). Also keep in mind that my commentary need not reflect Chomsky s own views, and I am responsible for mistakes in interpretation. Juan Uriagereka Part One [The original does not have parts, but I have divided it in two for presentation purposes.] What follows extends and revises an earlier paper ("Minimalist Inquiries," MI), which outlines a framework for pursuit of the so-called "minimalist program," one of a number of alternatives that are currently being explored. 1 The shared goal is to formulate in a clear and useful way -- and to the extent possible to answer -- a fundamental question of the study of language, which until recently could hardly be considered seriously, and may still be premature: To what extent is the human faculty of language FL an optimal solution to minimal design specifications, conditions that must be satisfied for language to be usable at all? This is a humble statement, to be kept in mind. We may think of these specifications as "legibility conditions": For each language L (a state of FL), the expressions generated by L must be "legible" to systems that access these objects at the interface between FL and external systems -- external to FL, internal to the person. The strongest minimalist thesis SMT would hold that language is an optimal solution to such conditions. SMT, or a weaker version, becomes an empirical thesis insofar as we are able to determine interface conditions and to clarify notions of "good design." This should be considered a goal of this paper. Keep in mind also the very next sentence... 1 Chomsky (1998). For background see references cited there and Lasnik (1999), among many others.

2 While SMT cannot be seriously entertained, there is by now reason to believe that in nontrivial respects some such thesis holds, a surprising conclusion insofar as it is true, with broad implications for the study of language, and well beyond. Note the indefinite article: an optimal solution. "Good design" conditions are in part a matter of empirical discovery, though within general guidelines of an a prioristic character, a familiar feature of rational inquiry. In the early days of the modern scientific revolution, for example, there was much concern about the interplay of experiment and mathematical reasoning in determining the nature of the world. Even the most extreme proponents of deductive reasoning from first principles, Descartes for example, held that experiment was critically necessary to discover which of the reasonable options was instantiated in the actual world. Similar issues arise in the case at hand. 2 In other words, you don t go to the math department to get your notion of good design, although you can ask them; but next you have to match that nice notion against your findings. Tenable or not, SMT sets an appropriate standard for true explanation: anything that falls short is to that extent descriptive, introducing mechanisms that would not be found in a "more perfect" system satisfying just legibility conditions. If empirical evidence requires mechanisms that are "imperfections," they call for some independent account: perhaps path-dependent evolutionary history, properties of the brain, or some other source. It is worthwhile to keep this standard of explanation in mind whether or not some version of a minimalist thesis turns out to be valid. This is a methodological point, of course. These considerations bear directly on parametric variation, in this case yielding conclusions that are familiar features of linguistic inquiry. Any such variation is a prima facie imperfection: one seeks to restrict the variety for this reason alone. Observe how this is an argument different from the familiar learnability one. The same goal is grounded in independent concerns of explanatory adequacy/learnability, which require further that ineliminable parameters be easily detectable in data available for language acquisition. Both kinds of considerations (related, though distinct) indicate that study of language should be guided by the uniformity principle (1): 2 See MI for discussion.

3 (1) In the absence of compelling evidence to the contrary, assume languages to be uniform, with variety restricted to easily detectable properties of utterances One familiar application is the thesis that basic inflectional properties are universal though phonetically manifested in various ways (or not at all), stimulated by Jean- Roger Vergnaud's influential Case-theoretic proposals 20 years ago. Another is the thesis proposed by Hagit Borer and others that parametric variation is restricted to the lexicon, and insofar as syntactic computation is concerned, to a narrow category of morphological properties, primarily inflectional. These have been highly productive guidelines for research, extending earlier efforts with similar motivation (e.g., efforts t o reduce the variety of phrase structure and transformational rules). What counts as "compelling" is, of course, a matter of judgment: there is no algorithm to determine when apparently disconfirming evidence is real or is the effect of unknown factors, hence to be held in abeyance. I d like to stress the continuity expressed in this paragraph, which could be extended to many other examples. In recent years a trend has emerged that sees minimalism as a break from the GB tradition. Although minimalism does use notions that were only incidental in GB (economy of derivations, economy of representations, last resort) the general concerns are old, and the technical solutions emerged in the course of limitations that previous versions of the model faced. On such grounds, we try to eliminate levels apart from the interface levels, Whatever those turn out to be, not an easy question. and to maintain a bare phrase structure theory and the inclusiveness condition, which bars introduction of new elements (features) in the course of computation: indices, traces, etc. The indispensable operation of a recursive system is Merge (or some variant of it), which takes two syntactic objects _ and _ and forms the new object _ = {_, _}. We assume further that _ is of some determinate type: it has label LB(_). In the best case, LB(_) = LB(_) or LB(_), determined by general algorithm. 3 This note is longer in the M-version, which speaks of how little use there is in this paper for labels, except for that of the new object formed (Collins s locus ). What follows is fast, and not without problems noted in my commentary to Minimalist Inquiries. In the M-version this paragraph is slightly elaborated on, but the technical issues persist. Let s set all of this to the side, though, assuming some appropriate, minimalist definition of command, e.g. in terms of active derivational workspaces. 3 On the possibility of dispensing with labels, see Collins (1999).

4 Merge yields two natural relations: Sister and Immediately-Contain (IC). Allowing ourselves the operation of transitive closure, we derive the relations Contain, Identity, and C-command. While Merge "comes free," any other operation requires justification. Similarly, any features of lexical items that are not interpretable at the interface require justification. That includes most (maybe all) phonological features; these must be deleted or converted to interface-interpretable form by the phonological component. One might ask to what extent the phonological component is an optimal solution to the requirement of relating syntactic input to legible form, a hard question, not yet seriously addressed. We keep here to narrow syntax: computation of LF. The empirical facts make it clear that there are (LF-)uninterpretable inflectional features that enter into agreement relations with interpretable inflectional features. Thus, the _-features of T (Tense) are uninterpretable and agree with the interpretable _-features of a nominal that may be local or remote, yielding the surface effect of noun-verb agreement. The M-version has a footnote 4 associated to Tense, concerning the possibility of Agr nodes (because of Alternative II below). Some obvious difficulties with this alternative are mentioned there as well. The obvious conclusion, which we adopt, is that the agreement relation removes the uninterpretable features from the narrow syntax, allowing derivations to converge at LF while remaining intact for the phonological component (with language-variant PFmanifestation). We therefore have a relation Agree holding between _ and _, where _ has interpretable inflectional features and _ has uninterpretable ones, which delete under Agree. An empirical question to ask here is whether this informal statement constitutes a necessary arrangement of interpretable and uninterpretable features, and if so what this follows from (good design?). In particular it is worth asking whether uninterpretable features can check uninterpretable ones, and if not why not (bad design?). See more on this below. The relation Agree and uninterpretable features are prima facie imperfections. In MI and earlier work it is suggested that both may be part of an optimal solution to minimal design specifications by virtue of their role in establishing the property of "displacement," which has (at least plausible) external motivation in terms of distinct kinds of semantic interpretation and perhaps processing.

5 The suggestion of an argument here is this: say there is some external motivation for displacement (unclear, but say that s true). Then the mechanism of uninterpretable features would be there in order to guarantee the displacement. There s a bit of a tautological flavor, though: since the system has displacement apparently associated to uninterpretable features, let s say that this is a consequence of good design, even if we don t understand how. An alternative, which as far as I can see has the same empirical coverage, would be to say: this is an imperfection, but one which the system uses to its advantage. Think of this as a viral theory: you have to get rid of uninterpretable features and you do, via movement. As a consequence the system finds itself in a new set of representations which can be used for new interpretive purposes. The second part has a reasoning like this. If so, displacement is only an apparent imperfection of natural language, as are the devices that implement it. If the alternative I mentioned is true, the imperfection is real. The facts remain unchanged, although not their motivation. Displacement is implemented by selecting a target and a related category to be moved to a position determined by the target. 4 The target also determines the kind of category that can be moved to this position. If uninterpretable inflectional features are the devices that implement displacement, we expect to find uninterpretable features of three kinds: This is the technical statement. Note that the position here is this: there are three things you need in order to implement displacement via uninterpretable features (those in (2) below) and indeed you find three elements involved: the featural set, the EPP property, and Case. (2) (i) to select a target _ [the probe] (ii) to determine whether _ offers a position for movement and if so, what kind of category can move to that position [in the M-version the and clause is moved to (i); this shows, I think, that good design is not always obvious...] (iii) to select the category _ that is moved That seems correct. For movement of a nominal to T, for example, the _-set and EPPfeature of _ = T serve the functions (i), (ii), respectively. 4 4 Terminology is often metaphoric here and below, adopted for expository convenience.

6 The M-version has a fn. 6 here: The EPP feature alone is not sufficient to identify a target; the _ -set (or comparable features, for other targets) is required to determine what kind of category K is sought. And see my comment regarding (ii) above... The category _ that is moved has uninterpretable structural Case, serving the function (iii). Agree is the relation between T and the moved category -- more precisely, their relevant subparts. 5 This should be relevant, for instance, for adjectives. Let us say that the uninterpretable features of _ and _ render them active, so that matching leads t o agreement. Locality conditions yield an intervention effect if probe _ matches inactive _ which is closer to _ than matching _, barring Agree(_, _). The picture seems to generalize over an interesting range. To the extent that this is true, uninterpretable features and the Agree relation are not true "imperfections," despite appearances. Uninterpretability of features -- say, of phonological features, _-features of T, or its EPP-feature -- is not "stipulated." The existence of these features is a question of fact: does L have these properties or not? If it does (as appears to be the case), we have to recognize the fact and seek to explain it: in the best case by showing that these are only apparent imperfections, part of an optimal solution to design specifications. Though motivated at the interface, interpretability of a feature is an inherent property that is accessible throughout the derivation. The phonological properties [±continuant], for example, are motivated only at the interface, but these "abstract" features are accessible throughout the derivation, which ultimately eliminates them in favor of narrow phonetic features interpretable at the interface. This is not totally clear, and it is hard to distinguish from an alternative whereby interpretability is purely at the interface, but the interface is scattered. Similarly, interpretability of _-features ([+] for N, [-] for T) is accessible throughout the derivation. For convergence, uninterpretable features must be deleted; in narrow syntax, we assume, by the operation Agree, establishing an agreement relation under appropriate conditions. Suppose that L has generated the syntactic object K with label LB(K). On minimalist assumptions, LB(K) is the only element of K that is immediately accessible to L, 5 There is presumably a similar but distinct agreement relation, concord, involving Merge alone.

7 This is a natural stipulation, which tells you that the information under the top tier of categorial representation is inaccessible to the system. I m willing to grant that on minimalist assumptions, but an assumption does not follow, and as such it might be wrong. so it must be the element that activates Agree, by virtue of its uninterpretable features: these constitute a probe that seeks a matching goal In the M-version another collection of features is added here. within the domain of LB(K). What is the relation Match? The optimal candidate is Identity; we therefore take Match to be Identity. There are two large paragraphs missing in the early version. The key idea in those paragraphs is presented later on in the present text: uninterpretable features (only them) enter the derivation without values. To avoid a confusing terminological issue, let s clearly separate a feature, a value, and a dimension. A feature is a valued dimension; for instance N is a dimension + is a value and +N is a feature (basically, a concrete property). So what we re being told is that you can have mere dimensions in the lexicon without any value, if they are uninterpretable. These elements get their values determined by Agree, which results in their being deleted from narrow syntax, otherwise you couldn t tell them apart from interpretable features at LF, as those are valued to start with; nonetheless these guys remain in the phonological component. As a consequence of this implementation, it is not featural identity that is at stake in Agree, but actually dimensional identity. Chomsky s speaks in terms of non-distinctness, which seems unnecessary. In other words, all that a probe cares about is seeking an identical dimension lower down in its domain (not an identical feature). We can thus keep to the more minimalist notion of identity (which is obviously more elegant than nondistinctness ), and the important suggestion is this: grammar is sensitive to featural dimensions, not their specific values. That seems like a deep fact. Another important technical paragraph states how Spell-out removes LF material which is uninterpretable and transfers the relevant object (WITH the uninterpretable stuff) to the phonological component. Fn. 8 of the M-version discusses the technical reason pointed out in MI of why this sort of system is necessary (overt syntax eliminates uninterpretable features, but they still have to have an effect on PF, thus the distinction in the Minimalist Program between deletion and erasure ). Technically this is somewhat curious, I think, in that you need two representations of the relevant object K: one which is sent intact to

8 PF, and one which is sent to LF without uninterpretable stuff. Note also that Spell-out must determine what is uninterpretable, which is trivial prior to Agree (no values). This is used to motivate the cyclic application of Spell-out, or relevant information would be lost. Note also that if Spell-out applies without values having been assigned, Chomsky wants the derivation to crash at the interface. Here again you get that strange look ahead relevant to not having the interface apply in a scattered fashion. You want to say that things crash at the interface even when, strictly, the interface as a level is something that the system doesn t reach until the end. Of course, Chomsky says that interpretability is a property that obtains throughout the derivation, even prior to reaching the interpretable components. I guess it s a bit like saying that you can violate an American law, being an American citizen, even when you re traveling abroad. You ll be tried upon your arrival in the US, but nonetheless you violated the law already in the place you committed the crime. Needless to say, as a consequence of this system there s no covert/overt distinction in the cycle. Chomsky also wants the phonological cycle to proceed in parallel, quite literally. He has to say that, as properties of relevant objects that don t make it to LF do make it to PF. One possibility that this suggests, of course, is that there are two entirely different systems at work: the syntactic one, which works with purely abstract categories, and then the phonological one, which access the syntactic one at the point of Spell-Out; that would eliminate part of the redundancy, but I ll return to this issue when Chomsky comments on Distributed Morphology. I will keep here largely to Case-agreement and related systems: _-features, structural Case, EPP, A-movement, and the core functional categories T, C, v (T = tense, C = complementizer, v a light verb that introduces verbal phrases). 6 This note is extended upon in the M-version, where the v* terminology is thus obliquely introduced. There is also a quick comment on Quirky Case, understood as inherent Case with a structural Case feature (I m not sure I understand). Within these systems probe and goal match if features are valued for the goal and unvalued for the probe. In the M-version there s an important added sentence: If _ features were valued for the probe, it would be inactive and could drive no operation; if they were unvalued for the goal, they would receive no values from the (unvalued) matching features of the probe. Note, however, that this still leaves the possibility (in principle) of mixed bags of features, with some valued and unvalued combinations. If correct, the analysis should generalize to other core syntactic processes. Some 6 For expository purposes, we take the nominal with structural Case to be N and use T and C as cover terms for a richer array of functional categories, as in MI.

9 extensions to wh-movement are suggested in MI, but tentatively, for reasons indicated. This is the easiest case of A'-movement, since there are grounds to believe that features of probe and goal are involved. In other cases (e.g., topicalization, VPfronting), postulation of features is much more stipulative; and throughout, questions arise about intermediate stages of successive-cyclic movement and island conditions. Another humble admission. Keep it in mind if you work on A -systems... Matching of probe-goal induces Agree, eliminating uninterpretable features that activate them. A number of questions arise; specifically, with regard to the theses (3): ( 3 ) (i) Probe and goal must both be active for Agree to apply (ii) _ must have a complete set of _-features (it must be _-complete) to delete uninterpretable features of the paired matching element _ Completeness is a very interesting notion. Computationally, we will see that only complete categories are validly engaged in operations. I think representationally, also, there s something to pursue with regards to the notion completeness. After all, there is an obvious correlation between complete categories and syntactic/semantic independency (e.g. compare complete T and partial T, for instance, or think of the difference between fully referential D s and indefinite or non-referential ones as complete D vs partial D; basically, the more features a category has the richer its semantic independence). Let us tentatively adopt both theses, returning to the matter. 7 For the Case-agreement systems, the uninterpretable features are _-features of the probe and structural Case of the goal N. _-features of N are interpretable; hence N is active only when it has structural Case. Once the Case value is determined, N no longer enters into agreement relations and is "frozen in place" (under (3i)). Structural Case is not a feature of the probes (T, v), but The sentence: it deletes under agreement if the probe is appropriate -- _-complete, assuming (3ii). is substituted in the M-version for: it is assigned a value under agreement, then removed by Spell-out from the narrow syntax. The value assigned depends on the probe: Nominative (NOM) for T, accusative 7 One question is whether (i) and (ii), if valid, follow from other properties of FL. That seems plausible, but there are many factors to be considered.

10 (ACC) for v (alternatively ergative-absolutive, with different conditions). The change of heart is understandable: mere deletion of Case would result in no Case realization, which might be fine for LF interpretation, but not for PF. So the idea is that the system somehow removes Case from the system, although it has not been valued in the same way as probing features are. The bottom line still is this: Case itself is not matched, but deletes under matching of _-features. So no matter how we cut it, Case is different from the other features. This poses a non-trivial conceptual question: could the system have worked optimally without this particular activation system? Note, in particular, that it is not totally obvious why the system should involve different Case values; one could imagine a system whereby goals get activated on the basis of a single Case specification. Possibly, though, there are other reasons why the system requires different Case values, as opposed to a single one (which so far as I can see would have been enough for activation, since Case values carry absolutely no interpretive consequence, unlike all other values). The following paragraph has been eliminated from the M-version: We take uninterpretable features to be unvalued, receiving their values only under Agree. That is natural, given that the values are redundant. It is also empirically motivated by intervention effects (see MI). Accordingly, Match is not strictly speaking identity, but nondistinctness: same feature, independently of value. Of course, this has now been incorporated before, as already noted. In some cases, an active element E is unable to inactivate a matched element by deleting its unvalued features. E is defective, differing in some respect from otherwise identical active elements that induce deletion. The M-version contains the clarification: The simplest way to express the distinction, requiring no new mechanisms or features, is in terms of (3ii): a nondefective probe is _ -complete, a defective one is not. The M-version mentions participle-object constructions here, which can manifest (partial) _ -feature agreement without Case assignment (the participle is defective). Then other cases: The most familiar illustrations are raising constructions and their ECM counterparts, as schematically in (4i), where _ is the matrix clause, _ is an infinitival with YP a verbal phrase (the case most relevant here), and P is the probe: T with raising verb (case (ii)),

11 v with ECM transitive verb (case (iii)) 8 : (4) (i) [ _ P [ _ [SUBJ [H YP]]]] ^ ^ MATRIX INFINITIVAL probe T (ii) (a) there are likely to be awarded several prizes (b) several prizes are likely to be awarded probe v (iii)(a) we expect there to be awarded several prizes (b) we expect several prizes to be awarded The Case-agreement properties of SUBJ [in (i)], and its overt location, are determined by properties of the matrix probe P, not internally to _. _ is a TP with defective head T DEF, which is unable to determine Case-agreement but has an EPP-feature, overtly manifested in (iii). Raising-ECM parallels give good reason to believe that the EPPfeature is manifested in (ii) as well, by trace of the matrix subject; preference of Merge over (more complex) Move gives a plausible reason for the surface distinction between SPEC-T DEF in (ii) and in (iii) (see MI). In the (a) cases, the EPP-feature of T DEF is satisfied by Merge of expletive; in (b) by raising of the direct object. The following sentence has been displaced in the M-version, as noted: The simplest way to distinguish T DEF from a full probe, requiring no new mechanisms or features, is in terms of (3ii): a full probe is _-complete, a defective one is not. At this point there are several ways to proceed. This is going to discuss the possibility that the EPP holds in infinitival clauses (the conventional approach) or not, the second alternative. Alternative (I), the conventional approach adopted in MI, follows the path just outlined. SPEC-T DEF of _ is filled either by Merge or Move, then associated to the higher probe. Examples (4ii, iiib) illustrate raising to subject, leaving reconstruction sites. 9 With 8 8 In English, the expected form "awarded several prizes" surfaces more naturally as "several prizes awarded." We return to the matter. 9 Reconstruction for trace of A-movement, as distinct from PRO, was discovered (to my knowledge) by Luigi Burzio, later published in Burzio (1986): e.g., such distinctions as "one interpreter each (was assigned t/*planned PRO to speak) to the visiting diplomats." For extensive review of these topics, see Fox (1999), Sauerland (1998). For a different approach to A-chains particularly, see Lasnik (forthcoming). The first half of this note is interesting, since Chomsky does not usually speak of reconstruction in these terms.

12 P = T+raising verb, SPEC-T DEF may apparently be manifested: e.g., in Icelandic defective infinitivals, possibly reflecting the TEC option (see MI). This comment is interesting, because Chomsky does not usually recognize reconstruction possibilities for A-movement; we ll discuss the footnote later on. T DEF matches SUBJ in some of its features (to implement raising) but not all (to preclude inactivation). If P and T DEF match SUBJ in the feature [person], then categories with this feature, and only these, can undergo raising (nominals but not adjectivals); on the simplest assumptions, T DEF has no other _-features. The If P... part is straighforward, but the...and Tdef part is less obvious. The then categories... consequence would follow even if only P matches SUBJ in person. Then there is not much to say about the _ - features of Tdef. Expletives too must have the feature [person], since they raise; and pure expletives of the there-type should have no other formal features, on the simplest assumptions. One has to be careful with these shoulds, which can be misinterpreted. After all, we do know that expletives differ cross-linguistically, in terms of whether they do or do not induce agreement, or a definiteness effect, for instance; it would seem that the way to capture those differences is in terms of features other than [person] (in the more accurate terminology, a [person] dimension, or unvalued feature [person]). In a framework that dispenses with categorial features, as is reasonable on minimalist grounds, [person] plays the role formerly assigned to [D] or [N] features. 10 When _-complete, T values and deletes structural Case for N. The _-set of N (which is always _-complete) both values and deletes the _-features of T (with or without movement). 10 On eliminating categorial features in favor of root structures with functional heads, see Marantz (1997), MI. This suggests an approach to syntactic computation along the lines explored by Murgia 1999 for ellipsis, which relates to the issue posed for the parallel PF and LF objects, mentioned in the text. On light verbs in particular, see among others Hale and Keyser (1993), Harley (1995). Functional categories lacking semantic features require complication of phrase structure theory (see MI), a departure from good design to be avoided unless forced. AGR elements, for example, should arouse skepticism. right, but then see note 14. What follows is also plausible. Similarly, D -- or at least one variant of D -- might be associated with referentiality in some sense, not just treated as an automatic marker of "nominal category"; nonreferential nominals (nonspecifics, quantified and predicate nominals, etc.) need not then be assigned automatic D (at least, this variant of D).

13 A slightly worrisome issue here is that, clearly, languages differ in whether their _ -set of N is overt. It may be that, for instance, the _ -set of English N is actually not complete, and thus for instance is incapable of licensing the null category one that it licenses in Spanish or Portuguese, where it might be complete. If that straightforward point were correct, we d need to slightly modify our assumptions about how the N set works. I m raising this just as a matter of principle. With defective probe, agreement is not manifested and Case of the matched goal is not assigned a value: raising T exhibits no agreement, and participles lack person; But they do agree in number and often gender, which raises a warning flag (we return to participles). neither determines the Case of matched N, which depends on a higher non-defective probe, T or v (see (18), below). Similar properties hold in other constructions, e.g., attributive adjectival/participial constructions ("[old/smashed] car," "a car [old enough t o buy/smashed into pieces]"). Whatever the correct analysis may be, these constructions involve a relation between N and the head of the predicate phrase; the complete _-set of N values and deletes the matched uninterpretable features of the predicate, but the partial _-set of the predicate does not value and delete Case in N, which still has to satisfy the Case Filter. Right, so two possibilities emerge: a) these are special, probing proceding in the absence of Case activation; or b) there is also some form of Case activation in these instances, albeit not the usual one that shows up in argumental positions. For example, it might be possible that null Case is involved in these instances, and more generally whenever number (not person) features are checked. That of course presupposes a system of null Case assignment. Again, I m raising these as possibilities that exist in principle, and do not seem particularly unreasonable. In the MI framework C is one-one associated with _-complete T (T COMP ): (5) C selects T COMP ; V selects T DEF That s surely a fact, but one whose cause we d like to understand (e.g., why doesn t Tcomp select Tdef? What would be wrong with that?) Control structures and finite clauses have the selectional relation C-T COMP, while raising clauses have the relation V-T DEF. Of course, it is slightly odd to say that the T of Control structures is as

14 complete as the T of regular complement sentences, if only because in the former you have obligatory consecutio temporum. The reasons in MI were largely theory-internal, having to do with Case-agreement and related systems. The conclusions are consistent with those reached on other grounds. The earliest straightforward evidence that raising clauses fall together with finite TP, and control clauses with CP, was provided by Luigi Rizzi, who observed that control clauses, like CPs, are phonetically isolable in ways that raising clauses are not, nor of course finite TP (stranding its complementizer). 11 [this footnote contains relevant data] These conclusions suggest a possible recasting of the account of defective elements, Alternative (II), with the invariant property (6): 12 I will not comment on this alternative seriously because it is not pursued thoroughly here, and in any case it faces some non-trivial difficulties (e.g. the one pointed out in fn. 14, which forces the postulation of separate Agr nodes) (6) C is _-complete; T is _-complete only when necessary One case in which T must be _-complete, it could be argued, is selection by _-complete _ with uninterpretable features (specifically, _ = C). The selectional property could then be formulated in terms of Match/Agree: the _-features of _ have to be deleted under Agree by T, which therefore must be _-complete (crash with failure of match is detectable at once); we return to some problems. It is tempting to associate EPP with _-completeness: C, and T selected by C, are _-complete, and therefore allow an EPPfeature; T DEF cannot have an EPP-feature. Accordingly, there is no internal raising t o SPEC-T DEF ; raising is "in one fell swoop" in such constructions as (4ii), and there are no intermediate reconstruction sites. 13 Case-agreement and EPP proceed as before, with T COMP. The symmetry of raising and ECM constructions suggests that the analysis should be extended to the latter as well. 14 Second Merge of first-merged object of V makes little sense. 11 IMPORTANT DATA:E.g., "it is to go home (every evening) that John prefers (*seems)"; Rizzi (1982). Other early evidence involved reconstruction effects (see note 9). Rizzi explained the distinction in terms of government, a notion not readily available in the framework here. 12 Some related ideas are developed by Pesetsky and Torrego (forthcoming). 13 See Epstein and Seely (1999) for proposals to this effect. And Castillo, Drury, and Grohmann (1999), who carefully argue for Alternative II. 14 This note is important. The extension is suggested by proposals of Koizumi (1995), Lasnik (1999, forthcoming), and Epstein and Seely (1999), differing from one another and from the account here, which is, furthermore, oversimplified.

15 The account should be restated, as in the sources cited, in terms of an AGR node selecting V; and by symmetry, selecting T (AGR selected by appropriate v and C). It is, then, AGR and not T/v that is the locus of _-features, Case, and EPP, in the version presented here. Since I will not be pursuing this course, in part for reasons discussed later, I will keep to the simplified exposition. Just as C comp selects T COMP, we might expect v COMP to select V comp The following that is, V with a complete set of _-features; it is, then, the _-features of V that enter into the Case-agreement system, parallel to T comp selected by C. The light verb v is _- complete in a construction with full argument structure: call it v*, transitive v or experiencer. 15 The analogy, then, is C-T comp, v*-v comp. is partly deleted and partly put into fn. 9 (M) in the M-version. Being _-complete, C must select T comp for its unvalued features to delete, and it allows an EPP-feature. For the same reasons, v* (being _-complete) must select V COMP for its (unvalued) _-features to delete, and it allows an EPP-feature. For both C and v, the selectional property reduces to Match/Agree. Unless selected by C or v*, T and V are defective (raising T, passive/unaccusative V, respectively). They do not enter into Case-agreement, and have no EPP-feature. When selected by C or v*, T and V are _-complete, entering into Case-agreement structures (with raising of associate or not, depending on optionality of the permitted EPP-feature and availability of alternatives to satisfy it). In a transitive construction, the object agrees with V and is assigned Accusative Case (raising t o SPEC-V if V has an EPP-feature). There is no internal raising to SPEC-T def in raising or ECM constructions. 16 Consider ECM constructions more closely [under laternative II]. The verbal phrase is: v-[v-tp]. If the light verb v is _-incomplete (passive), then V is defective, as is T V selected by V (in raising/ecm constructions). If v = v*, then V must be _- complete for convergence. But (6) does not require that T v be _-complete in this case, because the _-set of V is valued and deleted independently by the embedded subject. Therefore T remains defective, and the raising/ecm parallelism remains intact. The discussion suggests that T should be construed as a substantive rather than especially this part: 15 Only v was considered in MI, and I will largely keep to it for exposition below. Where assigned by V, not v, Case is inherent. Quirky Case largely falls under general Case-assignment principles if understood to be inherent Case with an additional structural Case feature associated with _-complete v* (as in MI). We return to examples, some problematic. 16 For T comp, the EPP-feature is apparently obligatory; for V comp, as well, for the sources cited in fn.14.

16 a functional category, falling together with N and V, perhaps others, a possibility that is neutral between Alternatives (I) and (II). We can regard T as the locus of tense/event structure (see note 6). The C-T relation is therefore analogous to the v*-v relation. This effectively ends discussion of alternative II (only further reference in notes). Observe how the comment about T being substantive is highlighted, since Chomsky wants to use that later on within the confines of alternative I. Alternative (I) takes the locus of Case-agreement/EPP to be T, v*; Alternative (II) takes it to be T, V. Let us refer to the two alternatives as LOCUS Tv* and LOCUS TV, focusing on their basic conceptual difference: the choice of category relevant to Caseagreement/EPP. The agreement issue noted in footnote 14 is also a serious conceptual difference. Much of what follows is neutral among these alternatives or variants that merit consideration as well. 17 I will continue with the conventional choice LOCUS Tv*, returning to others where appropriate, along with lingering problems. Suppose that the label LB(K) of K has an uninterpretable selectional feature (by definition, an EPP-feature), which requires Merge in SPEC of LB(K). That can be satisfied by Merge of an expletive, in which case long-distance agreement may hold between LB(K) and the goal. Alternatively, an active goal G determines a category PP(G) (Pied-Piping), which is merged in SPEC-LB(K), yielding the displacement property. The uninterpretable selectional feature has gone by many names in the last forty years. To call it selectional suggests a connection with theories of selection that does not seem obvious. Moreover, it is peculiar that this property should be satisfiable either by expletive Merge or by Move; a disjunction always hides a generalization being missed. (You may think that Move includes Merge, but that s given up by the end of the paper.) Also, by allowing this sort of situation in, the elegance of the Probe/goal situation is considerably weakened. So it is possible that this is a genuine imperfection, and an interesting one (Chomsky doesn t even try to say that this one is not an imperfection, at least not until now). You should perhaps keep in mind how similar expletive-associate relations are to case-np relations in languages where the case-marker is overtly associated to a full phrase. A question to ask is whether expletiveassociate dependencies might not be (one of) the origins of case systems 17 One variant might supplement (6) by taking T (and substantive categories generally, including V) to be always defective, so that the locus of Nominative Case and subject-verb agreement is C, not T (see reference of note 12 for related ideas). In the notation just used, this alternative falls under LOCUS Cv*.

17 (but see fn. 18). The combination of Agree/Pied-Pipe/Merge is the composite operation Move, preempted where possible by the simpler operations Merge and Agree. 18 FL specifies the features F that are available to fix each particular language L. The MI framework takes L to be a derivational procedure that maps F to {EXP}, where an expression EXP is a set of interface representations. As a first approximation, take EXP to be {PF, LF}, these being symbolic objects at the sensorimotor and conceptualintentional interfaces, respectively. Note: this is just a first approximation, probably a wrong one. We adopt the conventional (usually tacit) assumption that L makes a one-time selection [F L ] from F. These are the features that enter into L; others can be disregarded in use of L. There is a serious issue lurking here. Chomsky appears to be taking the view that the features of the syntax are much like those of phonology, a subset of those allowed by universal phonetics. However, we saw before that phonology probably does not have to obey the uniformity principle in (1), in part because its properties are readily available to the learner. The issue for syntactic properties is much less clear, though. Two possibilities arise. One is that if a language presents no evidence to the learner of the presence of a given syntactic category, then that category simply does not exist. For example, take noun classifiers in Japanese and in English, overt in the first instance. Does English have them? According to this view, no. However, as a consequence, Japanese and English LFs are going to be radically different, for example in the quantificational system. That is not necessarily a problem for learnability if the English system is the default one and you need positive evidence for the Japanese one. There is still an issue, learnability aside, as to whether this situation would be an imperfection, which would create two rather different systems of thought (for this view, see e.g. Gil (1987)). The other alternative is this: both Japanese and English have classifiers (Muromatsu (1998)), albeit they are covert in the latter. In this general view the problem is descriptive: under what circumstances does a language has an overt or a null classifier. The learnability issue is 18 It has occasionally been suggested that the EPP-feature of the probe P is a Case-assigning feature F C. If correct, that would scarcely change the picture: as before, F C is an uninterpretable selectional feature inducing Merge (sometimes Move), distinct from the features of P that enter into Agree. But it does not seem to be correct. There is good reason to believe that structural Case correlates with agreement, hence also long-distance agreement (without raising to SPEC), and accordingly that EPP is satisfied without Case-assigment (see (18) and discussion). If either is correct (both seem to be), Case-assignment and EPP are independent phenomena.

18 unchanged and the elegance question is now resolved. But note that in the second instance the set of actual and possible syntactic features is really identical, and the task of the grammarian is essentially to find out the set of universal syntactic categories. Assume further that L assembles [F L ] to lexical items LI of a lexicon LEX, the LIs then entering into computations as units. In the simplest case, LEX is a single collection, but empirical phenomena might call for "distribution" of LEX, with late insertion in the manner of Distributed Morphology (DM). 19 In any case, we can think of LEX as in principle "Bloomfieldian," a "list of exceptions" that provides just the information required to yield the interface outputs, and does so in the best way, with least redundancy and complication. This aside is in slight antagonism with the Bloomfieldian character. After all, why should the lexicon be exceptional or accidental? Presumably because it is historical, to a large extent. If so, though, why should it be yielding any optimality? What kind of optimality? More or less than in phonology more generally? In the simplest case, the entry LI is a once-and-for-all collection (perhaps structured) of (A) phonological, (B) semantic, and (C) formal features. The parenthetical comment perhaps structured is also a bit perplexing. If these features are indeed structured, what does that structure follow from? It must be either accidental (in which case it is possibly irrelevant) or else follow from some principled mechanism. But what is the nature of that pre-derivational mechanism? This is something to keep in mind, also, for those theories of theta-roles which take them to be features of lexical items. If they are, they should be a random collection, with no hierarchical or other structural properties. If on the other hand roles are ordered, somehow, we want to understand what is the principle that predicts that ordering, and what component of grammar is responsible for it. The features of (A) are accessed in the phonological component, ultimately yielding a PF-interface representation; those of (B) are interpreted at LF; and those of (C) are accessible in the course of the narrow-syntactic derivation. Language design is such that (B) and (C) intersect, and are disjoint from (A), In principle, either language design or an accident of evolution resulted in this property, assuming it is right. It is perhaps worth emphasizing that if we just blame whatever property we find on language design, without any 19 On the current state of DM, see Harley and Noyer (1999).

19 justification, the theory becomes circular. though there is some evidence, to which we return, that presence or absence of features of (A) might have an effect on narrow syntactic computation. These are a kind of stylistic rules discussed at the end of the paper, which pose a non-trivial complication of the model, although one that has very interesting potential consequences. It also seems that FL may retain something like (B)-(C) and the narrow syntax in which they enter while the phonological component is replaced by other means of sensiromotor access to narrow-syntactic derivations, as in sign language. There s a presupposition here that I m not sure one can directly grant; that in signed languages you literally replace whatever sensori-motor mechanism you had in spoken languages. That of course presuppose that the latter is prior, e.g. in evolutionary terms. So far as I know, this is a controversial claim. Bottom line: signed languages function just like spoken languages do, with a different kind of sensorimotor support. So sound PF cannot be essential to the language faculty, it has to be something more abstract. Of particular interest is the subset of (C) that is not in (B): uninterpretable formal features that appear, prima facie, to violate conditions of optimal design. For open classes, the optimal account is typically the simplest, for obvious learnability reasons: LI is a unitary collection, including the phonological matrix. LEX is distributed when departure from the simplest account is warranted in favor of late insertion, typically for inflectional elements and suppletion. So there is a distinction being drawn here between substantive and grammatical items. Note that the fact that the first type is a unitary collection doesn t entail that it has to be inserted early. It is perfectly possible (hard to determine empirically) that the unitary collection in question is inserted immediately before Spell-Out. Throughout, answers depend on predictability of phonetic outcome by general phonological principles that satisfy UG conditions, and in all cases, the simplest choice should win. For roots and highly predictable inflectional elements (say, English progressive), the distinctions between single-li and several independent contributions to LI (as in DM systems postulating universal late insertion ULI) seem to have little empirical content, Again, you don t care whether the rules determining the outcome take

20 place at a single point, in a separate phonological cycle, or they are distributed throughout the syntactic cycle. However, in less predictable instances... but they might, for example, when an idiosyncratic feature F of a root has syntactic effects. ULI then requires postulation of a redundant syntactic feature F' as a "place holder" in narrow syntax for F, with a stipulation that F' must be replaced under late insertion by a root with F (i.e., F' is effectively identical with F). 20 A unitary LEX avoids the redundancy and stipulation (and is preferable on conceptual grounds in any event). The substantive results of DM remain unchanged. If that comment seemed incomprehensible, don t worry; it is. It pertains to how you want to do your morphology in those instances where it matters, and whether you need to code it in terms of an abstract feature in the syntax that you replace upon matching (a stipulative matching in this instance). Since I know very little about how morphology should be done, I will not dwell on this now. Narrow syntax maps a selection of choices from LEX to LF; the phonological component, in contrast, has further access to [F L ]. I suppose that, in some instances, PF has to manipulate the items that the syntax has provided (e.g. when you get a verb and a clitic together in the syntax, and then you delete some consonant or change some vowel). That s fine, however... Like the extraction of [F L ] from F, these assumptions, largely conventional, reduce the computational burden for the procedure L while adding new conceptual apparatus. I m not sure in what sense what I said reduces computational complexity, to be honest. More controversially, MI extends the same reasoning to individual derivations: L makes a one-time selection of a lexical array LA, a collection of LIs (a "numeration" if some are selected more than once), and maps LA to EXP. Again, there is a reduction of computational burden, in this case a vast reduction, since LEX, which virtually exhausts L, need no longer be accessed in the derivation once LA is selected. That s certainly true, but then I m perplexed: why did the system take the step of reducing F to Fl, in turn assembled into LEX, if after all this is all nothing compared to the reduction from LEX to LA? Why not go from F to LA? Two answers come to mind. a) The intermediate steps (Fl, LEX) happen to be historical accidents or evolutionary ones; yes, it would have 20 For an illustration in the case of Latin deponent verbs, see Embick (1999).